Classifiers confidence in opencv face detector

Question

I'm using opencv's har cascade face detector (cv.HaarDetectObjects) in python.

for example:

    faces = cv.HaarDetectObjects(grayscale, cascade, storage, 1.2, 2,
    cv.CV_HAAR_DO_CANNY_PRUNING, (50,50))

       for f in faces:
           print(f)

This will print a list of detections in this form:

 ((174, 54, 114, 114), 53)
 ((22, 51, 121, 121), 36)
 ((321, 56, 114, 114), 21)
 ((173, 263, 125, 125), 51)
 ((323, 272, 114, 114), 20)
 ((26, 271, 121, 121), 36)

Where each line represent a detection. The first 4 numbers are the x,y location of the top-left point, and the height, width of the bounding box. The last number is (quoting from the openCV documentation) the number of neighbors.

I guess I have two questions:

1) What does the last number mean? I couldn't find any reference to that when googling.

2) (more important)Is there a way to get a confidence score for each detection? How much is the face classifier certain that the detection corresponds to a real face?

Thanks

Answer 1

1) The detection code produces more than one detection for an object - e.g. in different scales, slightly shifted, etc. The detections are then grouped and the number of neighbours in such a group is the number returned. See also Viola Jones paper, paragraph 5.6 (http://research.microsoft.com/en-us/um/people/viola/Pubs/Detect/violaJones_IJCV.pdf) and OpenCV source.

2) You can possibly use the number of neighbours as some measure of confidence.